[4H1-OS-2a] ニュースメディアのデータサイエンス(1/2)

[4H1-OS-2a-05] Generating Pairwise Learning Data for Click-through Rate Prediction of News Articles

〇Shotaro Ishihara¹, Yasufumi Nakama (1. Nikkei Inc.)

Keywords:pairwise learning, click-through rate prediction, randomized controlled trial, multimodal

In online news websites, the headlines and thumbnail images of articles are displayed in a list, and they are important navigation links to individual article pages. If we can predict the click-through rate (CTR) of readers to the article pages, we can assist the editors in creating article headlines and setting thumbnail images. However, the CTR that can be observed in the access log is heavily affected by the display position, and it is difficult to predict the CTR by machine learning using data on a single article. This paper proposes a method to construct a pairwise dataset based on the information such as similarity of data and display position, and create a model to predict the CTR in the framework of pairwise learning. In the experiment, we verified the usefulness of the proposed method by using the actual access log and discuss the potential of the practical use of editing support.